Вы находитесь на странице: 1из 26

3 SPATIAL VARIABILITY USING RANDOM FIELDS

3. 1 Need for spatial variability characterization in design


Many quantities such as properties of materials, concentrations of pollutants, loads etc in civil
engineering have spatial variations. Variations are expressed in terms of mean or average
values and the coefficients of variation defined in terms of the ratio of standard deviation and
mean value expressed as percentage. In addition, the distance over which the variations are
well correlated also plays a significant role.
A successful design depends largely on how best the designer selects the basic parameters of
the loading/site under consideration from in-situ and/or laboratory test results. Probabilistic
methods in civil engineering have received considerable attention in the recent years and the
incorporation of soil variability in civil/geotechnical designs has become important.
Considerable work was carried out in the area of geotechnical engineering. Guidelines such
as those of JCSS (2000) have also been developed in this context. Dasaka (2005) presented a
comprehensive compilation on spatial variability of soils. In the following sections, spatial
soil variability of soils is addressed and the concepts are applicable to any other property
variations as well. Soil has high variability compared to manufactured materials like steel or
cement, where variability in material properties is less, as they are produced under high
quality control.
3.2 Characterization of variability of design parameters
It is generally agreed that the variability associated with geotechnical properties should be
divided in to three main sources, viz., inherent variability, measurement uncertainty, and
transformation uncertainty (Baecher and Christian 2003; Ang and Tang 1984).
3.2.1 Inherent variability


The inherent variability of a soil parameter is attributed to the natural geological processes,
which are responsible for depositional behaviour and stress history of soil under
consideration. The fluctuations of soil property about the mean can be modelled using a zero-
mean stationary random field (Vanmarcke 1977). A detailed list of the fluctuations in terms
of coefficients of variation for some of the laboratory and in-situ soil parameters, along with
the respective scales of fluctuation in horizontal and vertical directions are presented in
Baecher and Christian (2003).
3.2.2 Measurement uncertainty
Measurement uncertainty is described in terms of accuracy and is affected by bias (systematic
error) and precision (random error). It arises mainly from three sources, viz., equipment
errors, procedural-operator errors, and random testing effects, and can be evaluated from data
provided by the manufacturer, operator responsible for laboratory tests and/or scaled tests.
Nonetheless the recommendations from regulatory authorities regarding the quality of
produced data, the measuring equipment and other devices responsible for the measurement
of in-situ or laboratory soil properties often show variations in its geometry, however small it
may be. There may be many limitations in the formulation of guidelines for testing, and the
understanding and implementation of these guidelines vary from operator to operator and
contribute to procedural-operator errors in the measurement. The third factor, which
contributes to the measurement uncertainty, random testing error, refers to the remaining
scatter in the test results that is not assignable to specific testing parameters and is not caused
by inherent soil variability.


3.2.3 Transformation uncertainty
9

Computation models, especially in the geotechnical field contain considerable uncertainties
due to various reasons, e.g. simplification of the equilibrium or deformation analysis,
ignoring 3-D effects etc. Expected mean values and standard deviations of these factors may
be assessed on the basis of empirical or experimental data, on comparison with more
advanced computation models. Many design parameters used in geotechnical engineering are
obtained from in-situ and laboratory test results. To account for this uncertainty, the model or
transformation uncertainty parameter is used.
3.3.4 Evaluation design parameter uncertainty
The total uncertainty of design parameter from the above three sources of uncertainty is
combined in a consistent and logical manner using a simple second-moment probabilistic
method. The design parameter may be represented as
( ) ,
m d
T = (1)
where
m
is the measured property of soil parameter obtained from either a laboratory or in-
situ test. The measured property can be represented in terms of algebraic sum of non-
stationary trend, t, stationary fluctuating component, w, and measurement uncertainty, e. is
the transformation uncertainty, which arises due to the uncertainty in transforming the in-situ
or laboratory measured soil property to the design parameter using a transformation equation
of the form shown in Equation 1. Hence, the design property can be represented by Equation
2.
( ) , e w t T
d
+ + = (2)
Phoon and Kulhawy (1999b) expressed the above equation in terms of Taylor series.
Linearizing the Taylor series after terminating the higher order terms at mean values of soil
10

parameters leads to the Equation 3 for soil design property, subsequently the mean and
variance of design property are expressed as given in Equations 4 and 5.
( )
( ) ( ) ( ) 0 , 0 , 0 ,
0 ,
t t t
d
T
e
T
e
w
T
w t T

+ (3)
( ) 0 , t T m
d

(4)
2
2
2
2
2
2
2

SD
T
SD
e
T
SD
w
T
SD
e w d

= (5)
The resulting variance of design parameter after incorporating the spatial average is given by
( )
2
2
2
2
2 2
2
2

SD
T
SD
e
T
SD L
w
T
SD
e w a

= (6)
Of the above, the treatment and evaluation of inherent soil variability assumes considerable
importance as the uncertainties from measurements and transformation process can be
handled if proper testing methods are adopted and transformation errors are quantified.
Approaches for evaluation of inherent soil variability are developed based on random fields
and a brief description of the theory and its relevance to characterisation of soil spatial
variability is described in the following sections.
3.4 Random field Theory
Soil properties exhibit an inherent spatial variation, i.e., its value changes from point to point.
Vanmarcke (1977a; 1983) provided a major contribution to the study of spatial variability of
geotechnical materials using random field theory. In order to describe a soil property
stochastically, Vanmarcke (1977a) stated that three parameters are needed to be described: (i)
the mean (ii) the standard deviation (or the variance, or the coefficient of variation); and (iii)
the scale of fluctuation. He introduced the new parameter, scale of fluctuation, which
11

accounts for the distance within which the soil property shows relatively strong correlation
from point-to-point.
Figure 3.1(a) shows a typical spatially variable soil profile showing the trend, fluctuating
component, and vertical scale of fluctuation. Small values of scale of fluctuation imply rapid
fluctuations about the mean, whereas large values suggest a slowly varying property, with
respect to the average.

(a) (b)

Figure 3.1(a). Definition of various statistical parameters of a soil property (Phoon and
Kulhawy 1999a); (b) approximate definition of the scale of fluctuation (Vanmarcke
1977a)
Vanmarcke (1977a) demonstrated a simple procedure to evaluate an approximate value of the
vertical scale of fluctuation, as shown in Figure 3.1(b), which shows that the scale of
fluctuation is related to the average distance between intersections, or crossings, of the soil
property and the mean.
A random field is a conceivable model to characterize continuous spatial fluctuations of a soil
property within a soil unit. In this concept, the actual value of a soil property at each location
within the unit is assumed to be a realization of a random variable. Usually, parameters of the
random field model have to be determined from only one realization. Therefore the random
12

field model should satisfy certain ergodicity conditions at least locally. If a time average does
not give complete representation of full ensemble, system is non-ergodic. The random field is
fully described by the autocovariance function, which can be estimated by fitting empirical
autocovariance data using a simple one-parameter theoretical model. This function is
commonly normalized by the variance to form the autocorrelation function. Conventionally,
the trend function is approximately removed by least square regression analysis. The
remaining fluctuating component, x(z), is then assumed to be a zero-mean stationary random
field. When the spacing between two sample points exceeds the scale of fluctuation, it can be
assumed that little correlation exists between the fluctuations in the measurements. Fenton
(1999a & b) observed that the scale of fluctuation often appears to increase with sampling
domain.
3.4.1 Statistical homogeneity
Statistical homogeneity in a strict sense means that the entire joint probability density
function (joint pdf) of soil property values at an arbitrary number of locations within the soil
unit is invariable under an arbitrary common translation of the locations. A more relaxed
criterion is that expected mean value and variance of the soil property is constant throughout
the soil unit and that the covariance of the soil property values at two locations is a function
of the separation distance. Random fields satisfying only the relaxed criteria are called
stationary in a weak sense.
Statistical homogeneity (or stationarity) of a data set is an important prerequisite for
statistical treatment of geotechnical data and subsequent analysis and design of foundations.
In physical sense, stationarity arises in soils, which are formed with similar material type and
under similar geological processes. Improper qualification of a soil profile in terms of the
statistical homogeneity leads to biased estimate of variance of the mean observation in the
13

soil data. The entire soil profile within the zone of influence is divided into number of
statistically homogeneous or stationary sections, and the data within each layer has to be
analysed separately for further statistical analysis. Hence, the partition of the soil profile into
stationary sections plays a crucial role in the evaluation of soil statistical parameters such as
variance.
3.4.2 Tests for statistical homogeneity
The methods available for statistical homogeneity are broadly categorised as parametric tests
and non-parametric tests. The parametric tests require assumptions about the underlying
population distribution. These tests give a precise picture about the stationarity (Phoon et al.
2003a).
In geostatistical literature, many classical tests for verification of stationarity have been
developed, such as Kendalls test, Statistical run test (Phoon et al. 2003a). Invariably, all
these classical tests are based on the important assumption that the data are independent.
When these tests are used to verify the spatially correlated data, a large amount of bias
appears in the evaluation of statistical parameters, and misleads the results of the analysis. To
overcome this deficiency, Kulathilake and Ghosh (1988), Kulathilake and Um (2003), and
Phoon et al. (2003a) proposed advanced methods to evaluate the statistical homogeneous
layers in a given soil profile. The method proposed by Kulathilake and Ghosh (1988),
Kulathilake and Um (2003) is semi-empirical window based method, and the method
proposed by Phoon et al. (2003a) is an extension of the Bartlett test.
3.4.2.1 Kendalls test
The Kendall statistic is frequently used to test whether a data set follows a trend.
Kendalls is based on the ranks of observations. The test statistic, which is also the
measure of association in the sample, is given by

14

2 / ) 1 n ( n
S

= (7)
where n is the number of (X,Y) observations. To obtain S, and consequently , the following
procedure is followed.

1. Arrange the observations (X


i
, Y
i
) in a column according to the magnitude of the Xs,
with the smallest X first, the second smallest second, and so on. Then the Xs are said
to be in natural order.
2. Compare each Y value, one at a time, with each Y value appearing below it. In
making these comparisons, it is said that a pair of Y values (a Y being compared and
the Y below it) is in natural order if the Y below is larger than the Y above.
Conversely, a pair or Y values is in reverse natural order if the Y below is smaller
than the Y above.
3. Let P be the number of pairs in natural order and Q the number of pairs in reverse
natural order.
4. S is equal to the difference between P and Q;
A total of
2
) 1 n ( n
2
n

=

possible comparisons of Y values can be made in this manner. If all


the Y pairs are in natural order, then
2
) 1 n ( n
P

= , Q=0,
2
) 1 n ( n
0
2
) 1 n ( n
S

=

= , and hence
1
2 / ) 1 n ( n
2 / ) 1 n ( n
=

= , indicating perfect direct correlation between the observations of X and Y.


On the other hand, if all the Y pairs are in reverse natural order, we have P=0,
2
) 1 n ( n
Q

= ,
2
) 1 n ( n
2
) 1 n ( n
0 S

=

= , and 1
2 / ) 1 n ( n
2 / ) 1 n ( n
=


= , indicating a perfect inverse
correlation between the X and Y observations.
15

Hence cannot be greater than +1 or smaller than -1, thus, can be taken as a relative
measure of the extent of the disagreement between the observed orders of the Y. The strength
of the correlation is indicated by the magnitude of the absolute value of .
3.4.2.2 Statistical run test
In this procedure, a run is defined as a sequence of identical observations that is followed and
preceded by a different observation or no observation at all. The number of runs that occur in
a sequence of observations gives an indication as to whether or not results are independent
random observations of the same random variable. In this the hypothesis of statistical
homogeneity, i.e., trend-free data, is tested at any desired level of significance, , by
comparing the observed runs to the interval between . Here, n=N/2, N being
the total number of data points within a soil record. If the observed number of runs falls
outside the interval, the hypothesis would be rejected at the level of significance.
Otherwise, the hypothesis would be accepted.
2 / ; 2 / 1 ; n n
r and r

For testing a soil record with run test, the soil record is first divided into number of sections,
and variance of the data in each section is computed separately. The computed variance in
each section is compared with the median of the variances in all sections, and the number of
runs (r) is obtained. The record is said to be stationary or statistically homogeneous at
significance level of , if the condition given below is satisfied.
2 / ; 2 / 1 ; n n
r r r <

(8)
3.4.2.3 Bartletts approach
The classical Bartlett test is one of the important tests, which examines the equality of two or
multiple variances of independent data sets. The following steps are involved in the Bartletts
test.
16

The sampling window is divided into two equal segments and sample variance ( ) is
calculated from the data within each segment separately. For the case of two sample
variances, , the Bartlett test statistic is calculated as
2
2
2
1
s or s
2
2
2
1
s and s
( )
( ) [ ]
2
2
2
1
2
log log log 2
1 30259 . 2
s s s
C
m
B
stat
+

= (9)
where m=number of data points used to evaluate . The total variance, s
2
2
2
1
s or s
2
, is defined as

2
2
2
2
1 2
s s
s
+
= (10)
The constant C is given by
( ) 1 2
1
1

+ =
m
C (11)
While choosing the segment length, it should be remembered that m10 (Lacasse and Nadim
1996). In this technique, the Bartlett statistic profile for the whole data within the zone of
influence is generated by moving sampling window over the soil profile under consideration.
In the continuous Bartlett statistic profile, the sections between the significant peaks are
treated as statistically homogeneous or stationary layers, and each layer is treated separately
for further analysis.
3.4.2.4 Modified Bartlett technique
Phoon et al. (2003a, 2004) developed the Modified Bartlett technique to test the condition of
null hypothesis of stationarity of variance for correlated profiles suggested by conventional
statistical tests such as Bartlett test, Kendalls test etc, and to decide whether to accept or
reject the null hypothesis of stationarity for the correlated case. The modified Bartlett test
statistic can also be used advantageously to identify the potentially stationary layers within a
soil profile. This procedure was formulated using a set of numerically simulated correlated
17

soil profiles covering all the possible ranges of autocorrelation functions applicable to soil. In
this procedure, the test statistic to reject the null hypothesis of stationarity is taken as the peak
value of Bartlett statistic profile. The critical value of modified Bartlett statistic is chosen at
5% significance level, which is calculated from simulated soil profiles using multiple
regression approach, following five different autocorrelation functions, viz., single
exponential, double exponential, triangular, cosine exponential, and second-order Markov.
The data within each layer between the peaks in the Bartlett statistic profile are checked for
existence of trend. A particular trend is decided comparing the correlation length obtained by
fitting a theoretical function to sample autocorrelation data. If the correlation lengths of two
trends of consecutive order are identical, it is not required to go for higher order detrending
process. However, it is suggested that no more than quadratic trend is generally required to be
removed to transform a non-stationary data set to stationary data set (Jaksa et al. 1999).
The following dimensionless factors are obtained from the data within each layer.
Number of data points in one scale of fluctuation,
z
k

=

(12)
Normalized sampling length,
k
n
z k
z n T
I =

= =

1
(13)
Normalized segment length,
k
m
z k
z m W
I =

= =

2
(14)
where is the scale of fluctuation evaluated, and n is the total of data points in a soil record
of T. The Bartlett statistic profile is computed from the sample variances computed in two
contiguous windows. Hence, the total soil record length, T, should be greater than 2W. To
ensure that m10, the normalized segment length should be chosen as I
2
=1 for k10 and I
2
=2
for 5k<10 (Phoon et al. 2003a).
18

Equations 15 and 16 show the typical results obtained from regression analysis for I
2
equals
to 1 and 2 respectively for the single exponential simulated profiles. Similar formulations
have also been developed for other commonly encountered autocorrelation functions and
reported in Phoon et al. (2003a).
BB
crit
=(0.23k+0.71) ln(I
1
)+0.91k+0.23 for I
2
=1 (15)
BB
crit
=(0.36k+0.66) ln(I
1
)+1.31k-1.77 for I
2
=2 (16)
A comparison is made between the peaks of the Bartlett statistic within each layer with B
crit
obtained from the respective layer. If B
max
<B
crit
, the layer can be treated as statistically
homogeneous and hence, accept the null hypothesis of stationarity. Otherwise, if B
max
>B
crit
,
reject the null hypothesis of stationarity, and treat the sections on either side of the peaks in
the Bartlett statistic profile as stationary and repeat the above steps and evaluate whether
these sections satisfy the null hypothesis of stationarity. However, while dividing the sections
on either side of the peaks in the Bartlett statistic profile, it should be checked for m10,
where m is the number of data points in a segment.
3.4.2.5 Dual-window based method
Kulathilake and Ghosh (1988) and Kulathilake and Um (2003) proposed a simple window
based method to verify the statistical homogeneity of the soil profile using cone tip resistance
data. In this method, a continuous profile of BC distance is generated by moving two
contiguous sub-windows throughout the cone tip resistance profile. The distance BC, whose
units are same as q
c
, is the difference of the means at the interface between two contiguous
windows. In this method it is verified whether the mean of the soil property is constant with
depth, which is a prerequisite to satisfy the weak stationarity. At first, the elevation of the
window is taken at a level that coincides with the level of first data point in the q
c
profile.
After evaluating the BC distance, the whole window is moved down at a shift each time. The
19

computed distance BC is noted each time at the elevation coinciding the centre of the
window (i.e., the intersection of two contiguous sub-windows). This length of sub-window is
selected based on the premise that at least 10 data points are available within the sub-window.
The data within the two sub-windows is treated separately, and checked for linear trend in the
data of 10 points. The reason behind verifying the data with only linear trend is that within
0.2 m profile, higher-order trends are rarely encountered. In addition, in normally
consolidated soils, the overburden stress follows a linear trend with depth. Kulathilake and
Um (2003) suggested that the demarcation between existence and non-existence of a linear
trend in the data be assumed at a determination coefficient (R
2
) of 0.9. It means that if the R
2

value of theoretical linear fit is greater than 0.9, then the data set is said to be having a
linearly trend in it, if not the mean value is said to be constant throughout the sub-window.
Hence, within a window length (i.e., two contiguous windows) there exist four sets of
possibility of trend in the mean values. They are
1. Constant trend in both the contiguous sub-windows
2. Constant trend in upper sub-window and a linear trend in the lower sub-window
3. Linear trend in the upper sub-window and constant trend in the lower sub-window,
and
4. Linear trend in both the contiguous sub-windows.
The above four sets possibilities of trend within the contiguous windows are shown in Figure
3.2. As the distance BC increases, the heterogeneity of the q
c
at the intersection between
two sub-sections increases.
20


3.4.3 Trend removal
Once the statistically homogeneous layers are identified within a soil profile, the individual
statistical layers are checked for the existence of trend, and the same is removed before
evaluating the variance and autocorrelation characteristics of the data. In general, all soil
properties exhibit a trend with depth. The deterministic trend in the vertical soil profile may
be attributed to overburden stress, confining pressure and stress history of soil under study.
Generally, a smooth curve can be fitted using the Ordinary Least Square (OLS) method,
except in special cases such as varved clays, where periodic trends are clearly visible (Phoon
et al. 2003a). In most of the studies, the trend line is simply estimated by regression analysis
using either linear or polynomial curve fittings.
Other methods have also been applied, such as normalization with respect to some important
physical variables, differencing technique, which is routinely used by statisticians for
transforming a non-stationary time series to a stationary one. The normalization method of
trend removal with respect to a physical quantity accounts for systematic physical effects on
the soil profiles. In general, the detrending process is not unique. Different trend removal
B
C C C C
D D D
D
B
B
B
R
2

f
o
r

l
i
n
e
a
r

f
i
t
>
0
.
9

R
2

f
o
r

l
i
n
e
a
r

f
i
t
<
0
.
9

R
2

f
o
r

l
i
n
e
a
r

f
i
t
<
0
.
9

R
2

f
o
r

l
i
n
e
a
r

f
i
t
<
0
.
9

Figure 3.2. Evaluation of BC distance in various possible combinations
A
A A A
21

procedures will in most cases result in different values of the random fluctuating components
and different shapes of the autocorrelation function.
Baecher (1987) commented that the selection of a particular trend function is a decision on
how much of the spatial variability in the measurements is treated as a deterministic function
of space (i.e., trend) and how much is treated statistically and modelled as random processes.
However, the detrending process cannot be entirely arbitrary. After all, the fluctuating
components remaining in the detrended soil records must be stationary for meaningful
statistical analyses to be undertaken on limited data points. Clearly, the chosen trend function
should be reasonable in view of this stationary constraint. The scale of fluctuation or
autocorrelation distance evaluated from the non-stationary data is always higher than the
corresponding stationary data. In other words, the trend removal invariably reduces the scale
of fluctuation of the soil properties. One of the simplest methods to evaluate whether a linear
or 2
nd
order polynomial trend is sufficient to be removed from the experimental data is to
calculate the scale of fluctuation for the above both detrended data. If the evaluated scales of
fluctuation are closer to each other, a detrending process using the lesser degree polynomial
is chosen. In the limit, the scale of fluctuation is zero when the entire profile is treated as a
trend with zero random variation (Phoon et al. 2003a).
If a trend is evident in the measurements, it should be decided whether or not it should be
removed before statistical analysis of a set of raw data. An observed trend that has no
physical or geological basis or is not predictable must not be removed prior to statistical
analysis, since it is a part of the uncertainty to be characterized (Fenton 1999b). After
selecting a proper trend function for the data, the residuals off the trend are calculated. Phoon
et al. (2004) pointed out that trend removal is a complex problem, and there is at present no
fully satisfactory solution to it. The identified trend in the data is removed by employing any
of the following three widely used detrending methods.
22

3.4.3.1 Decomposition technique
In this method the data set is divided into stationary random field and nonstationary trend, by
using the results obtained from either a non-parametric test or a parametric test discussed in
the last section. Initially a linear trend is selected and removed from the original data. The
linearly detrended data is tested for the weak stationarity. If the residuals off the linear trend
do not satisfy the stationarity hypothesis, the above procedure is repeated by choosing a
higher order polynomial. However, it is suggested that no more than quadratic trend is
normally sufficient to transform a non-stationary data set to stationary data set (Jaksa et al.
1999), and keep them fairly stationary, as complete removal of the trend in the data is rarely
achieved.
3.4.3.2 Normalization technique
Normalisation of the data set with respect to a dominant parameter, such as cone tip
resistance, q
c
, effective overburden pressure, , is also used in geotechnical engineering to
make the data trend free.
'
vo

3.4.3.3 Differencing technique


In this method, a nonstationary data set is made stationary by using first, second or higher
order differencing technique. This method of testing a time series is suggested by Bowerman
and O'Connell (1983), which is suitable for data containing no seasonal variations. According
to Bowerman and O'Connell (1983) if the sample autocorrelation function for experimental
data dies down fairly quickly, the original data set can be treated as stationary. However, if
the sample autocorrelation function dies down extremely slow, then the original data set can
be transformed to a stationary set by taking first or second difference of original data set.
However, the term fairly quickly is rather subjective and extensive judgment is involved in
it. Moreover, it is observed that if no seasonal variations exist in the data, no more than
23

second difference is rarely needed to transform a nonstationary data to stationary data (Jaksa
et al. 1999).
2.4.4 Estimation of autocorrelation
Available methods for estimating the sample autocorrelation functions differ in their
statistical properties such as the degree of bias, sampling variability, ease of use,
computational requirements, etc.. The methods that are commonly used for this purpose are
method of moments, Bartletts approach, method based on maximum likelihood principle,
Geostatistics, etc. However, the method of moments is the most common used to estimate
sample correlation function of soil properties.
3.4.4.1 Method of moments
A classical way of describing random functions is through the autocorrelation function,
(z). It is the coefficient of correlation between values of a random function at separation
of k. The spatial correlation of a soil property can be modelled as the sum of a trend
component and a residual term (Vanmarcke 1977a), as shown in Equation 2.17.
x=z+e 17)
where x is the measurement at a given location, z is the trend component, and e is the residual
(deviation about the trend). The residuals off the trend tend to exhibit spatial correlation. The
degree of spatial correlation among the residuals can be expressed through an auto-
covariance function.
( ) ( ) ( ) [ ] ( ) ( ) [ ]
j j i i
Z t Z P Z t Z P E k c = (18)
where k is the vector of separation distance between point i and j, E[.] is the expectation
operator, P(Z
i
) is the data taken at location i, and t(Z
i
) is the value of the trend at location i.
24

The normalized form of the autocovariance function given in Equation 19 is known as the
autocorrelation function.
(k)= c[k]/c[0] (19)
where c[0] is the autocovariance function at zero separation distance, which is nothing but
variance data.
It is not possible to evaluate c
k
nor
k
with any certainty, but only to estimate them from
samples obtained from a population. As a result, the sample autocovariance at lag k, , and
sample autocorrelation at lag k, r
*
k
c
k
, are generally evaluated. The sample autocorrelation
function (ACF) is the graph of r
k
for lags k=0,1,2, h, where h is the maximum number of
lags allowable. Generally, h is taken as a quarter of total number of data points in time
series analysis of geotechnical data (Box and Jenkins 1970; Lumb 1975a). Beyond this
number, the number of pairs contributing to the autocorrelation function diminishes and
produces unreliable results. The sample ACF at lag k, r
k
, is generally evaluated using
( )
( )( )
( )
( )

=
+



=
N
i
i
k N
i
k i i
k
X X
N
X X X X
k N
r
1
2
1
1
1
1
1
(20)
If no measurement error or noise is present, r becomes equal to 1 at a lag distance of zero.
Statistically homogeneous data are used to evaluate the sample autocorrelation functions.
The autocorrelation characteristics of soil properties can be characterized either by
autocorrelation distance, or scale of fluctuation, which is theoretically equal to the area under
the correlation function. The scale of fluctuation (or correlation radius) for one dimensional
real field is defined as shown in Equation 21 (Vanmarcke 1977a).
( )

=
0
2 d (21)
25

More generally, the scale of fluctuation is defined as the radius of an equivalent unit step
correlation function, i.e., ()=1 for and =0 for > , being the Euclidian lag (JCSS
2000). The autocorrelation distance (or scale of fluctuation) is evaluated from the sample
autocorrelation function using method of fitting or based on Bartlett limits, which are
described in the following sections.
3.4.4.1.1 Method of fitting
Analytical expressions are fitted to the sample autocorrelation functions using regression
analysis based on least square error approach. The least square error is generally
characterised by the determination coefficient of the fit. Frequently used single-parameter
theoretical auto-correlation functions are exponential, squared exponential, though models
such as triangular, second order auto-regressive, spherical, etc. are also not uncommon to fit
the sample autocorrelation data in geotechnical engineering. Some of these models are given
in Table 3.1.
Table 3.1. Theoretical autocorrelation functions used to determine the autocorrelation
distance and scale of fluctuation, (Jaksa et al. 1999)






Model
No.
Theoretical
autocorrelation
function
Autocorrelation function
Auto-
correlation
distance,
Scale of
fluctuation,

1 Triangular

a z for
a z for
a
z
z
0
1


a a
2
Single
exponential
( ) b z
z
/ exp =


b 2b
3
Double
exponential
( ) ( )
2
/ exp c z
z
=

c c
4
Second-order
Markov
( )


+ =

d
z
d z
z
1 / exp
d 4d
5
Cosine
exponential
( )

e
z
e z
z
cos / exp
e e
26

Table 3.1 shows the autocorrelation distance and corresponding scale of fluctuation for
theoretical autocorrelation functions. A small scale of fluctuation () implies rapid
fluctuations about the mean and vice versa. and a large reduction in variance over any failure
plane; this results in a small spread of the performance function. Conversely a large
means much longer variations about the mean and results in smaller reduction in variance
over a failure plane (Mostyn and Soo 1992).
3.4.4.1. 2 Bartlett limits
In the field of time series analysis, the most commonly used method to compute the
autocorrelation distance is by Bartletts approximation. In this method the computed scale of
fluctuation corresponds to two standard errors of the estimate, i.e., the lag distance at which
the positive Bartletts limits given by Equation 2.21, superimposed on the autocorrelation plot
crosses the autocorrelation function (Jaksa et al. 1999).
N
r
h
96 . 1
= (22)
The scale of fluctuation of cone tip resistance varies from site to site. Moreover, it also varies
with type of soil, as Jaksa et al. (2004) reports smaller scales of fluctuation in sands than
clays due to their nature of formation. Further, Fenton and Vanmarcke (1998) argue that the
scale of fluctuation depends largely on the geological processes of transport of raw materials,
layer deposition, and common weathering rather than on the actual property studied.
Nonetheless, DeGroot and Baecher (1993) observed that the scale of fluctuation is also
function of sampling interval on in-situ measured property.
3.4.5 Effect of anisotropy in correlation scales
Most soils in nature are usually anisotropic due to their mode of sedimentation and
consolidation that cause preferred particle orientations. There are generally two types of
27

anisotropy. Inherent or initial anisotropy manifests itself in the soil deposits as a result of
applied stresses at the time of formulation in the form of first-structure on a macroscopic
scale or as a fabric orientation on the microscopic scale. Stress or induced anisotropy arises
from changes in the effective stress state produced by subsequent loading history. This
anisotropy can cause the elastic, strength and compressibility parameters of the soil deposits
to vary with direction, and hence cannot be ignored.
The soil properties exhibit large variations and their directional behaviour is observed by
many researchers (Vanmarcke 1983; Jaksa et al. 1999; Phoon and Kulhawy 1999a; Griffiths
and Fenton 2000; Nobahar and Popescu 2002; Fenton and Griffiths 2003; Jaksa et al. 2004;
Sivakumar Babu and Mukesh 2004; and Uzielli et al. 2005; Wei et al. 2005). The
autocorrelation distances in vertical and horizontal directions are never the same, but in
general, differ by an order of magnitude, with horizontal scale of fluctuation being higher
than that in the vertical (Uzielli et al. 2005). Attempts have been made in the literature to
formulate autocorrelation models for 1, 2, and 3-dimensional soil space (Vanmarcke 1977a;
and Kulathilake and Miller 1987). The effect of anisotropy of soil properties on the bearing
capacity in a probabilistic framework has not been studied extensively in the literature. Many
times, due to economic feasibility, speed of exploration, availability of equipment and time
constraints vertical cone penetration data alone is obtained and used in the evaluation of
strength properties (Wei et al. 2005).
The autocovariance structure is called isotropic if the normalized autocovariance depends on
the Euclidian distances between field points only, instead of the axis directional distance
components, components, i.e.,
( ) ( )
2 2 2
, , z y x z y x + + = (23)
28

Isotropy implies that the autocorrelation function is invariant to orthonormal transformation
of the field coordinates. Also the autocorrelation structure may be partly isotropic, for
example with respect to horizontal field directions:
( ) ( ) z y x z y x + = , , ,
2 2
(24)
For complete anisotropy, the exponential correlation function in 3-D space is
( )

=
z y x
D
z
D
y
D
x
z y x exp , , (25)
If an isotropy in the horizontal direction is assumed, then the exponential correlation function
shown in Equation 2.25 is reduced to
( )

+
=
z h
D
z
D
y x
z y x
2 2
exp , , (26)
Similar theoretical autocorrelation functions in 3-D field for other distributions can also be
formulated on the similar lines shown above.
3.4.6 Spatial averaging
Parameters in geotechnical analyses usually refer to averages of a soil property over a sliding
surface or a rupture zone in an ultimate failure analysis or significantly strained volumes in a
deformation analysis. If the dimensions of such surfaces or volumes exceed the scales of
fluctuation of the soil property, spatial averaging of fluctuations is substantial. This implies
that the variance of an averaged soil property over a sliding surface or affected volume is
likely to be substantially less than the field variance, which is mainly based on small sample
tests (e.g. triaxial tests) or small affected volumes in insitu tests (JCSS 2002).
Because of the spatial variability of soil properties, encountering a sufficiently low strength to
induce failure in localized areas is more likely than such an encounter over the entire zone of
29

influence. Both the conventional analyses based on the factor of safety and the simplified
probabilistic analyses fail to address this issue of scale of failure. Over the depth interval Z
the spatial average soil property is given as

=
Z
dz Z u
Z
Z u ) (
1
) ( (27)
The spatial average of the soil property u(x,y,z) over a volume V is given in the same way as

=
V
v
dxdydz z y x u
V
u ) , , (
1
(28)
Averaging distance depends on the nature of the problem in hand. For design of shallow
foundations in shear criterion, this distance is equal to the extent of shear failure zone within
the soil mass (Cherubini 2000). This distance for shallow foundations in cohesionless soil
subjected to vertical loading is approximately taken as 2B below the base of footing in the
vertical direction and 3.5B from the centre of footing in the horizontal direction, where B is
the width of the footing.
3.4.7 Evaluation of variance reduction function
The combined effect of spatial correlation and spatial averaging of soil properties over the
failure domain are beneficially utilized to reduce the variance of the measured data within the
zone of interest. The derivation of the variance reduction functions in terms of spatial
correlation and spatial average is described in the following section. JCSS (2002) presents the
evaluation of variance reduction function by both exact approach and simplified approach.
3.4.7.1 Variance reduction for data in 1-D space
The variability of soil property u
i
from point to point is measured by standard deviation
i

and the standard deviation of the spatial average property u
Z

is by

.
The larger the length
30

(or the volume) over which the property is averaged, higher is the fluctuation of u
i
that tends
to cancel out in the process of spatial averaging. This causes reduction in standard deviation
as the size of the averaging length or volume increases, which is given by
( )
i
z
u
Z

= (29)
A simple relationship of the variance reduction function in terms of scale of fluctuation and
averaging distance is given in Equation 2.30 (Vanmarcke 1977a).
0 . 1 0 . 1 ) (
0 . 1 ) (
2
2
=
>

L
Z
L
Z
Z
(30)
The Equation 2.30 indicates that with decrease in scale of fluctuation and increase in
averaging distance, the value of variance reduction function reduces, which in turn reduces
standard deviation of the spatially averaged soil property. In other words, the more erratic the
variation (i.e., less correlated the soil property) of the soil property with distance and larger
the soil domain considered, larger will be the reduction in variability of the average property.
This phenomenon is a result of the increasing likelihood that unusually high property values
at some point will be balanced by low values at other point (Vanmarcke 1977a). However,
Vanmarcke (1983) emphasized that the variance reduction function (T) is related to the
autocorrelation function () as given by.
( ) ( )

=
T T
dt dt t t
T
T
0 0
2 1 2 1
2
1
(31)
which reduces to
( ) ( )

=
T
d
T T
T
0
1
2

(32)
31

From Equation 2.32, the variance reduction functions for triangular, exponential, and squared
exponential autocorrelation functions can be worked out as given in Equations 2.33 to 2.35,
respectively.
( )


=
a T for
T
a
T
a
a T for
a
T
T
3
1
3
1
(33)
( ) ( )

= b T
b
T
T
b
T / exp 1 2
2
(34)
( )

= 1 exp
2 2
d
T
d
T
E
d
T
T
d
T (35)
where a, b, d are referred to as the autocorrelation distances, T is the averaging length, the
distance over which the geotechnical properties are averaged over a failure surface, and E()
is the error function, which increases from 0 to 1 as its argument increases from 0 to . In
terms of standard Gaussian cumulative distribution function E(u)=2[F
U
(u)-0.5].
As the averaging length, T the variance reduction function, (T) 0. In other words, the
chances associated with failure of huge volume of soil are very rare. In addition, (T) is
inversely proportional to T at very large values of T.
The variance reduction factor for averaging in one, 2 or 3-D random field may be
approximated as given in Equations .
( ) ( )
( )
( )
n n
n
n
n n n
L L for
L L
L L for L L


=
=
K
K
K K
1
1
1 1
2
1
(36)
where n=1, 2, 3, and L
1
, L
2
and L
3
are the lengths over which averaging takes place and
1
,

2
,
3
are the correlation radii. In case of separable autocorrelation functions, i.e. which can
32

be written as a multiplication of factors for each of the dimensions of a 2- or 3-D surface or
volume, the total variance reduction factor can, for the 3-D case be written as:
( ) ( ) ( ) (
3
2
2
2
1
2
3 2 1
2
L L L L L L = ) (37)
Similar to the above, Vanmarcke (1977a) also proposed an approximate and simplified
resultant variance reduction factor in 2-D space as the product of individual variance
reduction factors in vertical and horizontal directions in terms of scale of fluctuation () and
spatial averaging distance (L) in the respective directions as shown in Equation (38).
2 2 2
h v A
= (29).

The above propositions have been used in the analysis of spatial variability of soils and the
influence of spatial variability in foundation design is presented in Sivakumar Babu et al
(2005) Dasaka et al (2005) and Sumanta Haldar and Sivakumar Babu (2006).


33